The cancer cell proteome and transcriptome predicts sensitivity to targeted and cytotoxic drugs

This study shows that the proteomic and transcriptomic states of cancer cells are more predictive of drug sensitivity than genomic markers for most drugs, both within and across tumor types.


February 27, 2019
We thank both reviewers for their constructive and valuable comments, which helped to improve our manuscript. We have carefully and extensively revised the manuscript and performed a large number of additional analyses. The most important changes/additions are: (1) Compare genomic/transcriptomic/proteomic data as predictor of drug sensitivity in eight additional tumor types using data from the Cancer Cell Line Encyclopaedia (CCLE). The new CCLE analysis substantiates that proteomic and transcriptomic data can be used to predict drug sensitivity for a variety of drugs, whereas genomic data predict drug sensitivity only for a narrow selection of drugs, in agreement with our in-house melanoma panel (2) Investigate "transferability" of drug sensitivity models between tissue types also in CCLE data (3) Compare different machine learning algorithms, namely elastic net, Lasso, PLS, PLS with multiple outputs, regression trees, and a maximum correlation method (4) Compare different strategies to reduce high-dimensional transcriptomic data both in our in-house melanoma panel and CCLE data (5) Interpret our results strictly conservatively, which is also reflected in the updated title of the manuscript.
Please find a point-by-point response below.
Reviewer #1: This manuscript is interesting from the point of view of trying to infer drug sensitivity/resistance from the activation status of signaling pathways operating in cancer cells. However, there are major issues with the work that preclude publication in its current form and a lot of work would be required in order to make it fit for publication.
The main issues are: # The authors are drawing a lot of bold and broad conclusions from very limited data.
As an example, the authors state: "Our results suggest that evaluation of cell signaling prior to drug treatment may be sufficient to predict tumor response". A second example:"...showing that signaling data can predict drug viability." The authors should not present their interpretation of the data as if they were biological facts because, unfortunately, the authors do not follow up their hypothesis with experimentation that would underscore the above notions. As much as this reviewer likes the concept, the conclusion is not backed by the data.
We agree that certain statements in our original submission were not carefully phrased. To address the reviewer's comment we now compare genomic/transcriptomic/proteomic data as predictor of drug sensitivity in eight additional tumor types using publicly available data in CCLE, to test if our findings in melanoma can be extended to other tumor types (see Figure 5). In addition, we now interpret our results strictly conservatively, which is also reflected in the updated title of the manuscript.
# The number of RNAs, proteins and p-proteins that were used in this study are very limited. This reviewer doubts that broad conclusions can be drawn from such a limited set. It is further very unclear why the authors only used 200 mRNA species (starting from 26,000) and the criteria for filtering appear arbitrary.
We fully agree that feature selection is a crucial part of machine learning. In the revised manuscript, we now systematically assess various strategies for reducing high-dimensional transcriptomic data before making drug sensitivity predictions, both using our in-house melanoma panel and CCLE data (see Supplemental Figures  S12+S21). An interesting outcome is that the PLS drug sensitivity predictions are rather insensitive to the precise selection of genes, presumably due to correlations between groups of genes.
# No comparison is provided as to how the author's method performs vs other established ones (such as elastic net analysis or some such).

See next comment
# The authors should also better explain why they used PLS to model drug effects. The argument of "p>n" is not a valid one. If they had used the entire drug sensitivity matrix as input, PLS would have been the natural choice. But because they used each individual drug as input, (penalized) linear regression, like LASSO or elastic net, would have been better choices. Overall, the chosen method complicates interpretability and thus it remains unclear how good the models really are and which may have been a reason why the authors did not attempt follow-up experiments to prove their point. To address the reviewer's comment, we now include a thorough comparison of different machine learning algorithms, namely elastic net, Lasso, PLS, PLS with multiple outputs, regression trees, and a maximum correlation method, which can be found in Supplemental Figure S6+S7 for proteomic and transcriptomic data respectively. We find that PLS performs similarly well as elastic net and Lasso in proteomic data, but better than these algorithms in transcriptomic data, and hence we decided to use PLS as our primary method. Elastic net, which is a generalisation of Lasso, interestingly reduced to Lasso for our dataset.
Reviewer #2: The authors generated (phospho)proteomics and mutation data as well as drug response data for 109 drugs on a panel of 43 melanoma cell lines. Multiple analyses were performed and many broad claims are made. These include 1) that 'proteomic state is more predictive of drug sensitivity than transcriptomic state'; 2) that tissue type is equally predictive of response in the pan cancer setting as molecular data that 3) a proteomics predictor trained on one cancer type (melanoma) can not be transferred to another cancer type (endometrial) and that 4) on drug proteomics measurements are not more predictive for response than data collected in the treatment naive setting for a specific drug.
Major comments 1. While the study looks at potentially interesting problems fairly strong claims are made as outlined above and these can not be fully supported by the data presented.
Here are some specifics: 1.1 The claim that proteomics is more predictive than mRNA can, so some degree be made based on the data in Panel 4b as there are more red cells (significant) in the proteomics column as in the mRNA column. However, there are only 27 drugs so adding a significance to this number will likely be hard. Second, each data type is predictive for a different set of drugs -admittedly there is also a large amount of overlap. This is a more accurate and more important message. It is quite unexpected that mRNA data is not predictive of Vemurafenib response. In addition, one might question the validity of such a comparison on melanoma which is so strongly driven by MAPK signaling. Finally, showing this in only melanoma is interesting, but the message would be far more relevant and strong if is demonstrated in at least two additional cancer types (where the deck is not so obviously stacked in favor of phospho-proteomics).
To address the reviewer's comment we now compare genomic/transcriptomic/proteomic data as predictor of drug sensitivity in eight additional tumor types using publicly available CCLE data, showing that transcriptomic and proteomic data yield more accurate drug sensitivity predictions than genomic data for most drugs, similar to our in-house melanoma panel results (see Figure 5). In addition, we now interpret our results strictly conservatively, which is also reflected in the updated title of the manuscript.
We agree that there is a substantial overlap between transcriptomic and proteomic drug sensitivity predictions. We now correlate the proteomic and transcriptomic datasets, see answer to Comment 2 below.
1.2 The authors show, for a single drug (Dacometinib), that the pre-treatment data is roughly equally predictive than the on-treatment data. Once again, this is an important hypothesis to investigate, but it is hardly convincing to show this for a single drug. This should be performed far more extensively to draw any meaningful conclusions. This is the type of result that gets cited out of context very easily. In this context, the LINCS project contains a far more exhaustive data set. In addition, while the performance may be similar, it would be interesting to know whether the posttreatment measurements show different phenomena than the pre-treatment data. For example, are the residuals correlated?
We agree that using one single drug is not enough to generalise the idea that posttreatment data is no more predictive then pre-treatment data. Additionally, when preparing this resubmission we became aware of that the Cobimetinib AUC predictions were sensitive to small perturbations in the input data, such as small changes in filtering settings or excluding individual cell lines from the analysis (see also answer to Comment 7). Since we are no longer confident in the results presented in this section, we have decided to completely remove it from the manuscript.
While the LINCS project has amazing data for a limited number of cell lines, our analysis requires data for a large number of cell lines of the same tumor type with matched drug sensitivity. We therefore believe that LINCS, in its current state, would not be the optimal route to compare post-treatment vs pre-treatment data. We would like to point out that upon submission, our proteomic dataset was, to the best of our knowledge, the largest available dataset for a cell line panel of a single tumor type with matched drug sensitivity data.
1.3 The comparison of the classifier on melanoma and endometrial cancer and the associated conclusion that classifiers can not be transferred from one cancer type to the next is of little value as it is a single case. Again, it is interesting to know, but based on these results one can hardly make a statement such as 'This confirms that signaling states are predictive of drug response, but that these, in general, are not conserved across tumor types'.
We agree that generalising this point requires additional data. To address the reviewer's comment we now investigate the transferability of drug sensitivity models for eight additional tumor types using both proteomic and transcriptomic data from CCLE ( Figure 5e).
2. The authors present a new data type (phospho-proteomics, albeit limited in size). It will be very useful to relate this data type to the existing data types (mutations and RNAseq). For example, chart the associations between mutation and RNAseq data and phospho-proteomics. Which variables are correlated? What is the uniqueness and redundancy between the different data types? This is important to know as this provides a better stage for comparing the predictive capacity for drug response.
To address the reviewer's comment, we now correlate the proteomic and transcriptomic datasets. More precisely, we use PLS to predict proteomic markers from transcriptomic data, and find that the majority of proteomic markers can be rather well predicted from transcriptomic data (Supplemental Figure S15). This could explain why there is a substantial overlap in predictability of drug sensitivity from either the proteomic or transcriptomic dataset.
3. In the comparison of the mRNA and proteomics there are a number of differences that could explain the differences in performance. One of these is dimensionality.
The authors do perform a selection of genes in specific pathways, but reducing the number of features in the mRNA expression dataset based on the variance in the features is a simple but necessary experiment. Since these data types will likely reveal the same biological phenomena employing different genes (MAPK activity is reflected by phosphorylation events in the pathway, while RNAseq will capture downstream expression programs) selecting the same genes/proteins is also an interesting approach but likely not the optimal one.
We fully agree that feature selection is a crucial part of machine learning. In the revised manuscript, we now systematically assess various strategies for reducing high-dimensional transcriptomic data before making drug sensitivity predictions, both using our in-house melanoma panel and CCLE data (see Supplemental Figures  S12+S21). An interesting outcome is that the PLS drug sensitivity predictions are rather insensitive to the precise selection of genes, presumably due to correlations between groups of genes.
4. The authors do perform integration of RNAseq and proteomics data, but in the most simplistic way possible -early integration. It would be interesting to explore a number of interesting alternatives, including intermediate or late integration. In this way the complementarity of the data types will become more evident. Reliable biomarkers will likely emerge from both data types.
To address the reviewer's comment, we now perform also late integration, where the proteomic and transcriptomic models are weighted together by their respective inverse RMS error after cross-validation. Late and early integration resulted in similar performance (Supplemental Figure S13).
5. The authors use 'mean viability' as measure of drug response. If the drug concentrations are equally spaced, this is the same as the Area Under the doseresponse Curve (AUC). If this is the case, the authors should rather refer to it as the AUC. If the concentrations are not equidistant, we suggest calculating the actual AUC and using this. Otherwise results should be repeated with the IC50.
The drug concentrations are equally spaced in log-space, hence we now follow the reviewer's advice and replace "mean viability" with "AUC" throughout the manuscript. We have also added a subsection in Methods describing the drug sensitivity screening experiments in greater detail (Line 372).
Minor comments: 6. Motivate why the Spearman correlation was used as an additional filter criterion in addition to Pearson correlation. Please also indicate how the R2 is computed.
We agree that mixing Pearson correlation and Spearman rank correlation was unfortunate, and we now corrected this by removing the use of Pearson correlation from the manuscript. We now also explicitly define R 2 (Line 99).
7. The authors excluded WM115 based on its outlier status in viability predictions using either proteomics or transcriptomics data. Is that a good reason? How many dose points were used for the drug response?
We originally excluded WM115 based on two observations: firstly, inclusion of WM115 led to unstable sensitivity predictions for Cobimentinib when using leaveone-out cross-validation, secondly WM115 was an outlier with respect to PCA analysis of the proteomics dataset. Based on the reviewer's comment we now thoroughly analysed the stability of our results, and found that replacing leave-oneout cross-validation with repeated (N≈100-1000) 10-fold cross-validation greatly improved the stability of our results, and consequently we do no longer exclude WM115. We gauge the stability of the drug sensitivity predictions by the standard deviation of the correlation coefficient between predicted and measured AUC (after cross-validation) when using different (random)  8. There is fairly broad consensus that Elastic net is preferable over pure lasso and this should be used as benchmark.
To address the reviewer's comment, we now benchmark elastic net vs Lasso (a special case of Elastic net), together with several other machine learning algorithms (Supplemental Figure S6+S7), and find that elastic net and Lasso have identical performance in our dataset, however, Lasso has the advantage of being less computationally expensive.
9. line 116: 'When we tested the AKT-inhibitor Ipatasertib (GDC-0068) we found that it inhibited the growth of cell lines in the p-AKT high cluster more effectively than in the PTEN high cluster (adjusted p = 0:0014)'. I take it the authors hadn't tested the AKT-inhibitor before? Maybe they could phrase it as: "Motivated by the strong clustering based on components of the PI3K pathway, we tested the AKT-inhibitor x in addition to the y other drugs in the primary screen." The AKT inhibitor was part of the original drug screen, however, it was filtered out because of its low coefficient of variation in AUC across the cell line panel, as is apparent from Figure 2e. We now state this explicitly in the manuscript (Line 125).
10. line 168: The authors state: 'To reduce the complexity of our PLS models we tested to limit the number of (phospho)proteins to 10'. Sentence does not read well and why was 10 selected?
We removed this part from the manuscript. Instead we now show how crossvalidated RMS error depends on the number of non-zero Lasso variables for two of the most predictive/interesting drugs, MMAE and Vemurafenib (Supplemental Figure  S8-S11). The smaller the number of non-zero variables, the easier it is to implement the "composite biomarker" in practice.
We have changed the wording accordingly throughout the manuscript.
12. line 231: 'Furthermore, sound effect sizes (rho = 0.5) were deemed insignificant due to small sample size.' rho is not an effect size. # We have changed the wording accordingly. 18. The following figures really look alike and it takes some effort (going back to the text, reading the legend) to see where they are different. A small addition to the figure would take this effort away.

See comment 14+15.
19. Supp figs S15-16: Indicate in the figure that this is per-cancer, and indicate that one is CCLE and the other one is CTRP.

See comment 14+15.
20. Supp figs S17-19: Clearly indicate that this is a pan-cancer analysis (it says pancancer data in the legend, but you can still do a per-cancer analysis on a pan-cancer data set) on CTRP (to distinguish it from 5a-c, where you should probably indicate that it is a pan-cancer and CCLE analysis).

Reviewer #1 Review
Report for Author: The following broad claims were made in the previous version of the manuscript: 1. that proteomic state is more predictive of drug sensitivity than transcriptomic state; 2. that tissue type is equally predictive of response in the pan cancer setting as molecular data 3. that a proteomics predictor trained on one cancer type (melanoma) can not be transferred to another cancer type (endometrial) and that 4. that on-drug proteomics measurements are not more predictive for response than data collected in the treatment naive setting for a specific drug. After carrying out the reviewer comments, the authors managed to disprove all but one (Claim 3) of the claims made in the previous version of the manuscript. While Claim 3 still holds in the sense that for only 5 of the 25 drugs a predictor can be transferred from melanoma to endometrium, the authors also show that predictors trained on all-but-one cancer type can (in most cases) be transferred to the left out cancer type. Clearly the training set size plays an important role here, as the authors demonstrate, so there is still more exploration to be done regarding transferability. Here is the list of conclusions in the revised manuscript, as stated in the abstract: 1. drug sensitivity models trained on transcriptomic or proteomic data 2. outperform genomic based models for most drugs; 3. drug sensitivity models can be transferred between tumor types; 4. transcriptomic/proteomic signals may be alternative biomarker candidates for the stratification of patients without known genomic markers. Unfortunately these claims/conclusions are not novel as 1 and 2 have both been demonstrated in Iorio et al, 2016 (except for the proteomic markers, which have been investigated in later publications) and 3 is a consequence of 1.
Taken together, the authors did a reasonable effort to address the comments. Unfortunately, as a consequence, the new results contained in the revised manuscript are not sufficiently novel to warrant publication in this journal. Below we only treat the major comments.
Major comments 1. While the study looks at potentially interesting problems fairly strong claims are made as outlined above and these can not be fully supported by the data presented. Here are some specifics: 1.1 The claim that proteomics is more predictive than mRNA can, so some degree be made based on the data in Panel 4b as there are more red cells (significant) in the proteomics column as in the mRNA column. However, there are only 27 drugs so adding a significance to this number will likely be hard. Second, each data type is predictive for a different set of drugs -admittedly there is also a large amount of overlap. This is a more accurate and more important message. It is quite unexpected that mRNA data is not predictive of Vemurafenib response. In addition, one might question the validity of such a comparison on melanoma which is so strongly driven by MAPK signaling. Finally, showing this in only melanoma is interesting, but the message would be far more relevant and strong if is demonstrated in at least two additional cancer types (where the deck is not so obviously stacked in favor of phospho-proteomics).
Author response: To address the reviewer's comment we now compare genomic/transcriptomic/proteomic data as predictor of drug sensitivity in eight additional tumor types using publicly available CCLE data, showing that transcriptomic and proteomic data yield more accurate drug sensitivity predictions than genomic data for most drugs, similar to our in-house melanoma panel results (see Figure 5). In addition, we now interpret our results strictly conservatively, which is also reflected in the updated title of the manuscript. We agree that there is a substantial overlap between transcriptomic and proteomic drug sensitivity predictions. We now correlate the proteomic and transcriptomic datasets, see answer to Comment 2 below.
Reviewer reply: The authors carried out the suggested experiments and these disprove their initial point. Now the conclusion is that proteomics and transcriptomics are equally predictive, which has also been demonstrated in, amongst other, Roumeliotis et al. Cell Reports, 2017.
1.2 The authors show, for a single drug (Dacometinib), that the pre-treatment data is roughly equally predictive than the on-treatment data. Once again, this is an important hypothesis to investigate, but it is hardly convincing to show this for a single drug. This should be performed far more extensively to draw any meaningful conclusions. This is the type of result that gets cited out of context very easily. In this context, the LINCS project contains a far more exhaustive data set. In addition, while the performance may be similar, it would be interesting to know whether the posttreatment measurements show different phenomena than the pre-treatment data. For example, are the residuals correlated?
Author response: We agree that using one single drug is not enough to generalise the idea that post-treatment data is no more predictive then pre-treatment data. Additionally, when preparing this resubmission we became aware of that the Cobimetinib AUC predictions were sensitive to small perturbations in the input data, such as small changes in filtering settings or excluding individual cell lines from the analysis (see also answer to Comment 7). Since we are no longer confident in the results presented in this section, we have decided to completely remove it from the manuscript. While the LINCS project has amazing data for a limited number of cell lines, our analysis requires data for a large number of cell lines of the same tumor type with matched drug sensitivity. We therefore believe that LINCS, in its current state, would not be the optimal route to compare post-treatment vs pre-treatment data. We would like to point out that upon submission, our proteomic dataset was, to the best of our knowledge, the largest available dataset for a cell line panel of a single tumor type with matched drug sensitivity data.
Reviewer reply: Not applicable anymore as authors discovered that the data are not robust enough to include in the manuscript. Regarding the novelty of the dataset: the mass spectrometry dataset published by Roumeliotis et al, Cell Report, 2107 is of comparable size in terms of the number of samples (n=50) but larger in terms of the number of features (thousands of phosphosites and proteins).

1.3
The comparison of the classifier on melanoma and endometrial cancer and the associated conclusion that classifiers can not be transferred from one cancer type to the next is of little value as it is a single case. Again, it is interesting to know, but based on these results one can hardly make a statement such as 'This confirms that signaling states are predictive of drug response, but that these, in general, are not conserved across tumor types'.
Author response: We agree that generalising this point requires additional data. To address the reviewer's comment we now investigate the transferability of drug sensitivity models for eight additional tumor types using both proteomic and transcriptomic data from CCLE (Figure 5e).
Reviewer reply: We appreciate the effort to thoroughly investigate this point. As only 5 of the 25 drug predictors can be transferred from melanoma to endometrial cancer we agree with the authors that only a 'certain degree cross-cancer predictability is possible'. In the cross transfer between all-but-one cancer type, there seems to be better transferability of a proteomic and transcriptomic classifier trained on all-but-one cancer type and applied to the left-out cancer type. This implies that there are pan-cancer biomarkers on the transcriptomic and proteomic level. A very similar analysis, albeit not in this leave-one-cancer-type out format and only for transcriptomics data, this has already been demonstrated in Iorio et al, 2016, Figure 5. This is, in my opinion, the only remaining novel finding in the manuscript, but is as such not sufficient for publication in this journal.
2. The authors present a new data type (phospho-proteomics, albeit limited in size). It will be very useful to relate this data type to the existing data types (mutations and RNAseq). For example, chart the associations between mutation and RNAseq data and phospho-proteomics. Which variables are correlated? What is the uniqueness and redundancy between the different data types? This is important to know as this provides a better stage for comparing the predictive capacity for drug response.
Author response: To address the reviewer's comment, we now correlate the proteomic and transcriptomic datasets. More precisely, we use PLS to predict proteomic markers from transcriptomic data, and find that the majority of proteomic markers can be rather well predicted from transcriptomic data (Supplemental Figure S15). This could explain why there is a substantial overlap in predictability of drug sensitivity from either the proteomic or transcriptomic dataset.
Reviewer Reply: For transcriptomics and proteomics, this is the expected outcome: there is a large degree of collinearity between transcriptomics and proteomics, which explains their comparable overall performance. However, we also suggested that the correlation between mutations and proteomics/transcriptomics be investigated to learn which drugs can be predicted by mutations and not by the other data types and vice versa. The author response only partially addresses our comment.
3. In the comparison of the mRNA and proteomics there are a number of differences that could explain the differences in performance. One of these is dimensionality. The authors do perform a selection of genes in specific pathways, but reducing the number of features in the mRNA expression dataset based on the variance in the features is a simple but necessary experiment. Since these data types will likely reveal the same biological phenomena employing different genes (MAPK activity is reflected by phosphorylation events in the pathway, while RNAseq will capture downstream expression programs) selecting the same genes/proteins is also an interesting approach but likely not the optimal one.
Author response: We fully agree that feature selection is a crucial part of machine learning. In the revised manuscript, we now systematically assess various strategies for reducing highdimensional transcriptomic data before making drug sensitivity predictions, both using our inhouse melanoma panel and CCLE data (see Supplemental Figures S12+S21). An interesting outcome is that the PLS drug sensitivity predictions are rather insensitive to the precise selection of genes, presumably due to correlations between groups of genes.
Reviewer reply: This is of academic interest as the new results (Comment 1) now show that there is no performance difference between proteomics and transcriptomics. The authors did show that proteomics and transcriptomics data show a large degree of correlation.
4. The authors do perform integration of RNAseq and proteomics data, but in the most simplistic way possible -early integration. It would be interesting to explore a number of interesting alternatives, including intermediate or late integration. In this way the complementarity of the data types will become more evident. Reliable biomarkers will likely emerge from both data types.
Author response: To address the reviewer's comment, we now perform also late integration, where the proteomic and transcriptomic models are weighted together by their respective inverse RMS error after cross-validation. Late and early integration resulted in similar performance (Supplemental Figure S13). Author response: The drug concentrations are equally spaced in log-space, hence we now follow the reviewer's advice and replace "mean viability" with "AUC" throughout the manuscript. We have also added a subsection in Methods describing the drug sensitivity screening experiments in greater detail (Line 372).

Reviewer #2 Review
Report for Author: The reviewed manuscript has addressed many but not all of the issues raised by the reviewers. Hence the manuscript has improved but still falls short of the expectations for a paper in this journal. Specifically, 1. This reviewer feels that far too much of the manuscript is devoted to the almost trivial case of BRAF. Not only is this well known, the authors make several statements in this section that are simply incorrect or confusing. The reason why only some melanoma lines respond to BRAF inhibitors is becase they are driven by the V600E mutation. Therefore, it is entirely clear that this genomic information predicts drug response. It is therefore trivial to conclude that "Our results show that the key driver mutation in melanoma, BRAF V600E/K , was only a strong predictor of drug sensitivity for drugs that target the mutated molecule itself." Similarly, the title of the next section states an entirely trivial fact for the same reason. The V600E is what is driving the tumor. Not the other mutions in exons.
In this regard it is also confusing that the authors state in the introduction that "This often renders cross-entity biomarkers too unspecific to be used to stratify patients". If a melanoma patient presents with a V600E mutationi in BRAF, the patient will almost certainly be treated with a BRAF inhibitor.
Next, the statement "Again, only the BRAF V600E/K dependent drugs Vemurafenib and Dabrafenib could be predicted ignificantly..." is plain wrong. Both drugs inhibit BRAF wt just as much as the V600E mutant. It is a common misconceptionn that these BRAF inhibitors are specific to the mutant protein. They are not. the only reason why these drugs work in BRAF V600E patients is because the mutant protein drives the tumor.
Another main criticism to the initial manuscript was that none of the new hypothesis was validated by independent experiments. This was not addressed in the revision. Thank you for transferring your revised manuscript entitled "Signaling and expression states of cancer cells predict sensitivity to targeted and cytotoxic drugs" to Life Science Alliance.
Your work was reviewed at another journal twice before, and the editors transferred those reviewer reports and your response to us with your permission. The reviewers appreciated the technical quality of your work and the thorough revision performed, but thought that the conceptual advance provided remained somewhat limited. This is not a concern for us, and we would thus be happy to publish your paper in Life Science Alliance pending final revisions necessary to meet our formatting guidelines: -please provide a final point-by-point response to the remaining concerns of the reviewers, including references for drug specificity towards BRAFV600E/K -please upload all main figures as individual files, the legends should remain in the manuscript word docx file -please add a callout to Fig 2E in the manuscript text -please provide information in the Data availability section on NGS data accession codes -I would like to suggest to change the title slightly. How about: "The cancer cell proteome and transcriptome predicts sensitivity to targeted and cytotoxic drugs" If you are planning a press release on your work, please inform us immediately to allow informing our production team and scheduling a release date.
To upload the final version of your manuscript, please log in to your account: https://lsa.msubmit.net/cgi-bin/main.plex You will be guided to complete the submission of your revised manuscript and to fill in all necessary information. Please get in touch in case you do not know or remember your login name.
To avoid unnecessary delays in the acceptance and publication of your paper, please read the following information carefully.
A. FINAL FILES: These items are required for acceptance.
--An editable version of the final text (.DOC or .DOCX) is needed for copyediting (no PDFs).
--High-resolution figure, supplementary figure and video files uploaded as individual files: See our detailed guidelines for preparing your production-ready images, http://www.life-sciencealliance.org/authors --Summary blurb (enter in submission system): A short text summarizing in a single sentence the study (max. 200 characters including spaces). This text is used in conjunction with the titles of papers, hence should be informative and complementary to the title. It should describe the context and significance of the findings for a general readership; it should be written in the present tense and refer to the work in the third person. Author names should not be mentioned.

B. MANUSCRIPT ORGANIZATION AND FORMATTING:
Full guidelines are available on our Instructions for Authors page, http://www.life-sciencealliance.org/authors We encourage our authors to provide original source data, particularly uncropped/-processed electrophoretic blots and spreadsheets for the main figures of the manuscript. If you would like to add source data, we would welcome one PDF/Excel-file per figure for this information. These files will be linked online as supplementary "Source Data" files. **Submission of a paper that does not conform to Life Science Alliance guidelines will delay the acceptance of your manuscript.** **It is Life Science Alliance policy that if requested, original data images must be made available to the editors. Failure to provide original images upon request will result in unavoidable delays in publication. Please ensure that you have access to all original data images prior to final submission.** **The license to publish form must be signed before your manuscript can be sent to production. A link to the electronic license to publish form will be sent to the corresponding author only. Please take a moment to check your funder requirements.** **Reviews, decision letters, and point-by-point responses associated with peer-review at Life Science Alliance will be published online, alongside the manuscript. If you do want to opt out of this transparent process, please let us know immediately.** Thank you for your attention to these final processing requirements. Please revise and format the manuscript and upload materials within 7 days.
Thank you for this interesting contribution, we look forward to publishing your paper in Life Science Alliance.